最近更新: 4h ago

Instruction 基准测试

Instruction Following Evaluation 排行榜

Tests whether models follow explicit, verifiable constraints like 'write in more than 400 words' or 'mention AI at least 3 times'. All instructions have objectively verifiable criteria.

为什么重要： Measures instruction-following precision, critical for production applications. Models that score well here are more reliable in structured tasks.

顶级模型

93.5%

Gemini 3 Pro

平均评分

85.1%

共45个模型

已测试模型

指标: prompt-level accuracy

人类基准

评分范围: 0%–100%

IFEval Scores - Top 25 Models

Ranked by IFEval score (%)

LMMarketCap.com

模型排名

All models with a reported IFEval score, ranked by highest prompt-level accuracy.

排名模型评分性能

Gemini 3 Pro Google

93.5%

94%

93.5%

GPT-5.4 OpenAI

93.5%

94%

93.5%

GPT-5.2 OpenAI

93%

Claude Opus 4.6 Anthropic

92.8%

93%

92.8%

GPT-5.1 OpenAI

92.5%

93%

92.5%

Claude 3.7 Sonnet Anthropic

92.3%

92%

92.3%

Llama 3.3 70B Meta

92.1%

92%

92.1%

o3 OpenAI

92%

GPT-5 OpenAI

92%

#10

Claude Opus 4.5 Anthropic

91.5%

92%

91.5%

#11

Claude Sonnet 4.6 Anthropic

91%

#11

Grok 4 xAI

91%

#11

Gemini 3 Flash Google

91%

#14

Claude Sonnet 4 Anthropic

90.8%

91%

90.8%

#15

Claude Opus 4 Anthropic

90.5%

91%

90.5%

#16

o3-mini OpenAI

90.2%

90%

90.2%

#16

Claude Sonnet 4.5 Anthropic

90.2%

90%

90.2%

#18

o4-mini OpenAI

90%

#19

DeepSeek V3 (March 2025)DeepSeek

89%

#20

GPT-4.1 OpenAI

88.2%

88%

88.2%

#21

Claude 3.5 Sonnet Anthropic

88.1%

88%

88.1%

#22

Llama 4 Maverick Meta

88%

#22

Grok 3 xAI

88%

#24

Llama 3.1 405B Meta

87.5%

88%

87.5%

#25

Gemini 2.5 Pro Google

87.2%

87%

87.2%

#26

DeepSeek V3 DeepSeek

87.1%

87%

87.1%

#27

o1 OpenAI

86.5%

87%

86.5%

#27

Mistral Large 2 Mistral AI

86.5%

87%

86.5%

#29

Gemini 2.5 Flash Google

85.5%

86%

85.5%

#29

DeepSeek R1-0528 DeepSeek

85.5%

86%

85.5%

#31

GPT-4o OpenAI

84.3%

84%

84.3%

#32

Claude Haiku 4.5 Anthropic

84%

#33

Llama 3.1 70B Meta

83.6%

84%

83.6%

#34

Qwen 2.5 72B Alibaba

83.5%

84%

83.5%

#35

DeepSeek R1 DeepSeek

83.3%

83%

83.3%

#36

Gemini 2.0 Flash Google

82%

#37

GPT-4o mini OpenAI

80.4%

80%

80.4%

#38

Phi-4 Microsoft

80.1%

80%

80.1%

#39

Command R7B (12-2024)Cohere

77.1%

77%

77.1%

#40

Command R+Cohere

76.6%

77%

76.6%

#41

Qwen2.5 7B Instruct Alibaba

75.8%

76%

75.8%

#42

Qwen 2.5 Coder 32B Alibaba

72.7%

73%

72.7%

#43

Llama 3.1 8B Instruct Meta

72.0%

72%

72.0%

#44

Llama 3.2 3B Instruct Meta

68.5%

68%

68.5%

#45

Runway Gen-3 Alpha Runway

28.0%

28%

28.0%

关于 IFEval

全名: Instruction Following Evaluation
类别: Instruction
指标: prompt-level accuracy (%)
评分范围: 0%–100%
人类基准: 尚未确定
状态: 启用

Frequently Asked Questions

IFEval is a standardized evaluation that measures AI model performance on specific tasks. It provides comparable scores across different models, helping developers choose the right model for their needs.

Gemini 3 Pro currently holds the top score on the IFEval benchmark. See our full rankings table above for the complete leaderboard with 45 models.

We update benchmark data from multiple sources including HuggingFace open-source model leaderboards and LMArena. Scores are refreshed regularly as new evaluations are published and new models are released.

No. While IFEval is an important indicator, real-world performance depends on many factors including pricing, latency, context window, and specific task requirements. We recommend using our composite score which weighs multiple benchmarks and practical factors.

Instruction Following Evaluation 排行榜

模型排名

关于 IFEval

相关基准测试